[Sonnet 4.6 / Opus 4.7] Unexpected token consumption increase with `medium` effort after recent update

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Title

[Sonnet 4.6 / Opus 4.7] Unexpected token consumption increase with medium effort after recent update

---

Type of Behavior Issue

Unexpected / inconsistent resource usage (token consumption significantly higher than expected)

---

What You Asked Claude to Do

I used Claude Code in an iterative coding workflow with Sonnet 4.6 and medium effort.

Typical usage pattern:

• Asking for code improvements
• Iterating on functions
• Requesting explanations and refinements

Example command:

claude "refactor this function and improve readability"

No unusual prompt size or complexity compared to previous sessions.

---

What Claude Actually Did

1. Claude responded correctly to prompts
2. However, token consumption increased significantly compared to previous usage
3. Tokens were depleted much faster within the same type of workflow
4. The increase does not appear proportional to:

• prompt size
• response length
• task complexity

This results in much shorter usable sessions.

---

Expected Behavior

• Token consumption should remain relatively consistent for similar workflows
• medium effort should not behave significantly more expensive than before
• Changes in token usage should be predictable and documented

---

Files Affected

N/A (issue is related to model behavior and token usage, not file operations)

---

Permission Mode

Default Claude Code permissions (no special restrictions)

---

Can You Reproduce This?

Yes, consistently

---

Steps to Reproduce

1. Use Claude Code with Sonnet 4.6
2. Set effort to medium
3. Run iterative coding prompts (refactor, explain, improve code)
4. Observe token usage over multiple prompts
5. Compare with previous sessions or lower effort levels

---

Claude Model

claude-sonnet-4.6

---

Relevant Conversation

Not a single prompt issue — observed across multiple interactions.

Behavior pattern:

• Normal responses
• No visible increase in response size
• But significantly higher token consumption

---

Impact

High

• Tokens deplete much faster than before
• Reduces usability of Max plan
• Increases cost and unpredictability in workflows
• Affects production-like iterative usage

---

Claude Code Version

2.1.121

---

Platform

Claude Code (local CLI, macOS / darwin)

What Claude Actually Did

1. Claude responded correctly to all prompts (no functional errors observed)

1. However, token consumption per interaction appears significantly higher than in previous sessions using the same workflow

1. During iterative usage (multiple short coding/refactoring prompts), tokens are depleted much faster than expected

1. The increase in token usage does not seem proportional to:

• input prompt size
• response length
• task complexity

1. No visible change in response verbosity or quality that would justify the higher token usage

1. As a result, overall session capacity is reduced and tokens run out noticeably faster than before

1. This behavior started after the most recent update and is consistently reproducible

Expected Behavior

1. Token consumption should scale proportionally with:

• input prompt size
• response length
• task complexity

1. Similar workflows (iterative coding, short prompts) should result in comparable token usage across sessions

1. Using medium effort should not result in a significant or sudden increase in token consumption compared to previous versions

1. Token usage should remain predictable and stable for the same usage patterns over time

1. Any major changes in token consumption behavior should be clearly documented or communicated

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.121

Platform

Anthropic API

Additional Context

_No response_